Automatic Classification of Agricultural Crops Using Sentinel-2 Data in the Rainfed Zone of Southern Kazakhstan

Abstract

Satellite monitoring of agricultural crops plays a crucial role in ensuring food security and in the sustainable management of agricultural resources, particularly in regions dominated by rainfed farming, such as the Turkestan region of Kazakhstan. Many satellite monitoring tasks rely on remote identification of different types of cultivated crops. In developing the proposed method, we accounted for the temporal characteristics of crop growth and development in various climatic zones of rainfed agriculture, analyzed the dynamics of the Normalized Difference Vegetation Index (NDVI) together with ground-based data, and identified effective time periods and patterns for successful crop recognition. This study aims to develop and comparatively assess two methods for the automatic identification of cultivated crops in rainfed zones using Sentinel-2 satellite data for the years 2018 and 2022. The first method is based on detailed classification of pre-digitized field boundaries, providing high accuracy in satellite-based mapping. The second method represents a fully automated approach applied to large rainfed areas, emphasizing operational efficiency and scalability. The results obtained from both methods were validated against official national statistics, ground-based field surveys, and farm-level data. The findings indicate that the field-boundary-based method delivers significantly higher accuracy (average accuracy of 91.1%). While the automated rainfed-zone approach demonstrates lower accuracy (78%), it still produces acceptable results for large-scale monitoring, confirming its suitability for rapid assessment of sown areas. This research highlights the trade-off between the accuracy achieved through detailed field boundary digitization and the efficiency provided by an automated, scalable approach, offering valuable tools for agricultural production management.

1. Introduction

Agriculture serves as a cornerstone of food security and economic stability in many countries around the world, particularly in Central Asia, including Kazakhstan. In semi-arid regions such as the Turkestan region of Kazakhstan, the agricultural sector plays a crucial role in ensuring the supply of essential food products and supporting rural communities [1]. However, the reliance on rainfed agriculture, which depends entirely on natural precipitation, makes this sector particularly vulnerable to intensifying challenges associated with global climate change, such as droughts, extreme temperatures, and unpredictable rainfall patterns [2]. In addition to climatic factors, soil degradation, limited water resources, and inefficient land use also present significant obstacles to the sustainable development of agriculture under these conditions.

In this context, accurate and timely mapping of agricultural crops becomes not merely desirable but essential for optimizing land use, effectively monitoring the productivity of agricultural lands, responding promptly to crisis situations, and—critically—for the development and implementation of science-based, sustainable agricultural strategies at both regional and national levels [3,4,5,6,7,8]. Traditional methods of crop area assessment, such as ground-based field surveys, are extremely labor-intensive, financially costly, and resource-demanding, while also being limited in both scale and frequency. These limitations make them insufficient for delivering timely and comprehensive analysis over large territories. As a result, Earth observation technologies from space offer unprecedented opportunities for large-scale, systematic, and cost-effective analysis of agricultural lands [9,10].

In recent years, satellite data—particularly from the European Space Agency’s Sentinel-2 missions—have become one of the most widely used information sources for agricultural land monitoring, due to their unique combination of high spatial (10 m) and temporal (5–10 days under favorable conditions) resolution [8,9]. These features enable detailed monitoring of vegetation dynamics at the field level and facilitate the detection of key crop development stages. The Normalized Difference Vegetation Index (NDVI), as one of the most commonly used spectral indices, is widely applied to track phenological phases of crops such as emergence, active growth, peak biomass accumulation, and the subsequent senescence associated with crop maturity [11,12,13,14,15]. Its effectiveness is based on the reflectance differences between healthy vegetation in the red and near-infrared portions of the spectrum.

One of the main challenges in crop classification is the existence of species with similar phenological profiles, which complicates differentiation based solely on NDVI dynamics. For example, spring cereals and safflower often have overlapping growing seasons under semi-arid conditions [16,17,18,19,20,21]. To address this issue, additional indicators were employed. Notably, the Plowed Land Index (PLI), based on the timing of soil tillage activities, enables effective differentiation between winter and spring crops according to the timing of autumn and spring plowing, respectively [22,23].

Despite the considerable success of Sentinel-2 data applications, cloud cover in semi-arid regions—especially during peak vegetation periods—can significantly limit the availability of optical imagery. This necessitates the use of advanced preprocessing techniques such as compositing or integration with radar data [24].

In the SKT region, rainfed agriculture occupies a dominant position, encompassing a diverse array of cultivated crops, including winter cereals (primarily wheat), spring cereals (barley), oilseed crops such as safflower, and various perennial forage species like alfalfa [25,26,27,28]. Although general crop classification methods using satellite data are available, comprehensive approaches that take into account local agro-technical practices and the region’s unique climatic characteristics remain limited.

The present study aims to fill this gap by developing and rigorously evaluating two distinct methodologies for automated crop classification in rainfed areas of the South Kazakhstan Turkestan region. Both methods utilize Sentinel-2 satellite data, as well as vegetation and agrotechnical indicators such as NDVI dynamics during the growing season and the Plowed Land Index (PLI). These are implemented within the ArcGIS Pro 3.0.0 software environment using a mask of agricultural zones [29].

The first methodology focuses on classification within digitized field boundaries, ensuring high accuracy by excluding non-agricultural areas. The second methodology represents an automated approach applied over a broader rainfed agricultural zone, emphasizing scalability and operational efficiency [30].

The primary objective of this work is to develop a method for satellite-based identification of cultivated crop types based on temporal patterns in NDVI vegetation curves (including the onset of growth, peak biomass accumulation, and crop maturation) and agrotechnical indicators. A comparative validation of the classification algorithms is conducted to assess their effectiveness for rainfed agriculture monitoring [31,32,33].

The key research tasks include:

• Developing an automated and reproducible workflow for both methodologies;
• Conducting multi-level validation of the classification results using independent field data and official government statistics;
• Assessing the applicability of each methodology for solving practical challenges in regional agricultural management.

2. Materials and Methods

2.1. Study Area

The territory of the Turkestan Region (Southern Kazakhstan) is characterized by considerable diversity in topography, soils, and agro-climatic resources. Approximately half of the region’s area consists of plains and sandy landscapes, which are predominantly found in the northern and western parts. The Karatau mountain range, stretching from the northwest to the southeast, serves as a natural boundary between the plains and mountainous zones; most of its peaks reach elevations of around 1800 m above sea level. In the southeastern part of the region, there are low- and high-altitude ridges of the Western Tien Shan, with elevations ranging from 1500 to 2000 m to 2500–3500 m and above. This topographic variability has led to a pronounced heterogeneity in soil and climatic conditions, which, in turn, determines the agricultural specialization of different districts within the region [34].

According to statistical data for 2022, arable land in the SKT region covers 872.5 thousand hectares. Of this total, forage crops (primarily alfalfa) account for 229.7 thousand ha (26.3%), winter cereals occupy 228.1 thousand ha (26.1%), oilseed crops cover 71.3 thousand ha (8.2%), spring cereals make up 100.5 thousand ha (11.5%), vegetables and melons represent 131.7 thousand ha (15.1%), and cotton is cultivated on 106.4 thousand ha (12.2%), with the remaining area dedicated to other crops. Consequently, cereals, oilseeds, and forage crops constitute the majority of the sown area—629.6 thousand ha, or 72.1%. Crop areas fluctuate from year to year depending on weather conditions and crop rotation patterns.

The rainfed agricultural zone is primarily located in the foothill areas in the eastern part of the region and in the central plains. This zone is mainly used for the cultivation of winter cereals (wheat), forage crops (alfalfa), oilseeds (safflower), and spring cereals (barley) (Figure 1).

The irrigated agricultural zone is located in the central and southern parts along the Syr Darya River, as well as along the Arys River and the Arys-Turkestan canal. In these areas, crops such as cotton, maize, alfalfa, vegetables, and melons are cultivated under irrigation.

As part of this study, in addition to the satellite-based classification of crop areas in the rainfed zones of the Turkestan region, field survey points related to pastures were also used, covering the entire territory of the region. Based on these data, a pastureland map was developed, enabling the integration of vegetation information for both cultivated crops and natural forage lands (Figure 2).

The linkage of pasturelands with rainfed crop areas facilitates spatial analysis of land use structure, identification of transitional and mixed-use zones, and assessment of the degree of land resource competition between crop production and livestock grazing. This is particularly important in areas characterized by irregular precipitation and seasonal variability in productivity [35].

By overlaying the maps of field crop boundaries and pasture territories, it is possible to refine the delineation of functional land use, classify land parcels based on actual land utilization, and more accurately assess changes within the rainfed zone. This approach also improves the accuracy of satellite-based crop classification by excluding pasture areas from analysis or, alternatively, treating them as a separate land use class.

Further analysis focuses on identifying typical spectral and phenological characteristics of pasturelands in comparison with rainfed croplands. These insights can be leveraged to enhance the accuracy of automated land use classification based on RS data.

The agro-climatic characteristics of the study area reflect variations in thermal resources (sum of active temperatures above 10 °C (ΣT > 10 °C)) and moisture availability (Selianinov’s Hydrothermal Coefficient—HTC) during the growing season. Based on thermal and moisture conditions [1], the territory of the region can be divided into four agro-climatic zones:

• 1a—Very dry and hot agro-climatic zone (located in the northern part of the region up to the Karatau Mountains)
• 1b—Extremely dry and very hot agro-climatic zone (covers the Syr Darya valley, the right-bank Pre-Syrdarya plain, and the Kyzylkum desert)
• 2—Dry foothill zone (located in the central and eastern parts of the Karatau Mountains)
• 3—Mountain agro-climatic zone (includes the highest elevations of the Karatau range and Western Tien Shan) (Figure 3)

Zones 2 and 3 have higher moisture availability during the growing season and lower thermal resources. Therefore, less heat-demanding crops are cultivated in these areas, primarily under rainfed conditions and partially under irrigation (such as cereals, perennial grasses, oilseeds, safflower, fruit crops, etc.). These zones are also suitable for livestock farming.

During the growing season, the foothill areas of the region experience lower accumulation of thermal resources and increased moisture availability compared to the plains. Summer precipitation typically ranges from 50 to 100 mm in the foothills, while in the plains it is only 20 to 40 mm. The Hydrothermal Coefficient (HTC) of Selianinov for the warm period ranges from 0.5 to 0.7 and higher in the foothills, while in the plains it is about 0.2–0.3 [36,37].

These conditions create a more favorable environment for the vegetation of cereal, forage, and oilseed crops in the foothill zones due to the milder and more humid climate. However, because of insufficient warmth in the early spring, the vegetation of winter cereals and other crops in the foothills resumes slightly later than in the plains, typically in the first ten days of April. Consequently, the maturity and harvest of crops in these areas also occur later, generally during the second half of July (

Table 1. Agro-climatic characteristics by subregions of the Turkestan region.

Agroclimatic Region	1a	1b	II	III
District name	Very dry hot	Very dry and very hot	Dry foothill	Mountainous
HTC	0.1–0.3	0.1–0.3	0.3–0.5	>0.5
ΣT > 10 °C	3600–3900	4000–4600	3300–4400	<3300

).

In the plains of the SKT region, the onset of spring is marked by the earliest crossing of the mean daily air temperature above +5 °C. This results in an earlier start of spring sowing and vegetation of agricultural crops, typically occurring between March and early April, depending on weather conditions. The growth and development of crops in the plains usually take place under extremely hot and arid conditions, which suppress crop development and accelerate maturation. Consequently, harvesting often begins as early as late June—nearly a month earlier than in the foothill zones of the SKT region.

2.2. RS and Ground Data

To address the objectives of mapping rainfed agricultural crops in the SKT region, RS data were utilized as the basis for vegetation analysis and crop classification. The primary data source consisted of multispectral imagery from the Sentinel-2 satellite.

Sentinel-2 Level-2A imagery was selected due to its temporal (5 days) and spatial (10 m) resolutions. Unlike Level-1C data, which provide Top-of-Atmosphere (TOA) reflectance values and require additional atmospheric correction, Level-2A products are atmospherically corrected to Bottom-of-Atmosphere (BOA) reflectance. This means that distortions caused by atmospheric scattering and absorption (e.g., from aerosols and water vapor) have been removed. Additionally, Level-2A data undergo geometric correction and are already projected into the Universal Transverse Mercator (UTM) coordinate system using a digital elevation model, which ensures high geolocation accuracy and consistency across multi-temporal scenes [38]. These features are critical for deriving accurate vegetation indices, such as NDVI, and for conducting reliable temporal analyses of phenological crop changes.

The territory of the SKT region is covered by multiple Sentinel-2 tiles (coverage scheme shown in Figure 4). However, for the purposes of this study, which focuses on rainfed farming areas mainly located in the foothill and central parts of the region, only the following tiles were processed: 42TVP, 42WP, 42VN, 42WN, 42XN, 42XM, 42WM, 42VM, and 42VL. These tiles encompass the core rainfed zones and provide sufficient spatial coverage for the research [39,40].

The satellite data acquisition and processing period included autumn 2021 (October–November) and the 2022 growing season (March–July). This time frame was chosen to capture the complete phenological cycle of the major rainfed crops. Autumn imagery is essential for detecting the sowing and early development of winter crops, while spring and summer imagery are used to monitor the active growth, peak vegetation, and maturation of both winter and spring crops, including safflower, spring cereals, and perennial grasses. Additionally, satellite imagery from the 2018 growing season was also used for comparative purposes.

To support the analysis, calibration, and validation of the satellite data processing results, a set of ground-based field survey data was used. Field surveys were conducted between 2 and 7 June 2022, covering various agricultural fields with a total of 349 test GPS points collected along the survey routes. The survey results included crop type identification, precise GPS coordinates of observations, photographic documentation, and information on crop development stages and overall condition, including records of autumn and spring tillage operations.

Additionally, data from a previous field campaign conducted in 2018 were utilized during the development of the classification methodology. A map of the field survey routes and the locations of test farms in the SKT region is presented in Figure 5.

Alongside the field surveys, valuable data were collected in collaboration with local farms including “Bratya,” “Makulbek,” “Kyzylzhar,” and “Karabau.” The farmers from these operations provided detailed information on land use, crop types, sowing and harvesting dates, which served as an important resource for calibrating and validating the crop classification algorithm.

2.3. Methods

The preprocessing of RS data is a critically important step in ensuring the accuracy of subsequent analysis and classification of cultivated crops. All data processing procedures were conducted using ArcGIS Pro version 3.0.0, with the Spatial Analyst extension, which provides a comprehensive set of tools for raster data processing and analysis [41,42,43].

The first step involved the removal of clouds and associated shadows, which can significantly distort surface spectral reflectance values. To accomplish this, standard methods based on the Scene Classification Layer (SCL) mask provided with Sentinel-2 Level-2A data were employed [44]. This mask contains information about pixels classified as clouds, cloud shadows, water, and other non-informative features. Pixels identified as clouds or cloud shadows were excluded from further analysis (Figure 6).

Next, composite NDVI rasters were generated for each 3–5-day interval by averaging the NDVI values from all available cloud-free images within each time window. This approach minimizes the impact of residual atmospheric distortions and random noise, producing more stable and representative vegetation profiles for each pixel.

In addition to NDVI, the Plowed Land Index (PLI) was calculated to assist in crop differentiation. PLI was computed for two key periods: autumn (mid-September to early November 2021), when soil preparation for winter crops takes place, and spring (March–April 2022), when tillage for spring crops is performed. The PLI is based on analyzing changes in soil brightness, primarily in the Short-Wave Infrared (SWIR) spectral range, which is sensitive to soil moisture content and surface structure after tillage. The specific formula or method for calculating PLI is described in [25,45] and was subsequently adapted to local climatic conditions to effectively detect plowed areas within the Turkestan region.

At the final stage of preprocessing, a rainfed agriculture mask (Agri_zone.tif) was applied. This mask, developed from spatial data on the boundaries of rainfed agricultural areas in the Turkestan region, was used to exclude all non-agricultural lands (e.g., urban settlements, forests, water bodies) from the analysis, allowing the classification to focus solely on rainfed fields. The overall workflow for mapping cultivated crops under rainfed conditions in the Turkestan region is illustrated in Figure 7.

2.4. Classification Methodology

The classification of rainfed crops was conducted using an integrated approach that combines the analysis of vegetation curves derived from NDVI time series and agro-technical indicators obtained from the PLI. The entire classification process was implemented in ArcGIS Pro 3.0.0 using the Raster Calculator and a rule-based approach [46].

This approach is based on constructing a hierarchical system of logical conditions that differentiate crop types based on their unique spectral-temporal and agro-technical characteristics. Each rule is a combination of NDVI threshold values at key phenological stages and PLI data indicating the presence or absence of autumn/spring plowing.

For instance, the identification of winter cereals involved conditions reflecting autumn plowing with no evidence of spring tillage. Conversely, spring crops were identified by the presence of spring tillage and absence of autumn preparation. Perennial grasses were recognized by their consistently high NDVI values throughout the season without characteristic tillage patterns [47,48,49,50,51,52].

Two key types of NDVI composites were used:

• NDVI temporal composites: mean NDVI rasters calculated from all available cloud-free images within a 3–5-day window (e.g., NDVI_25apr_composite). These were created using the Cell Statistics—Mean function and used to track vegetation dynamics.
• NDVI peak composite (NDVI_peak): a raster reflecting the maximum NDVI value for each pixel over a characteristic peak-growth period for a specific crop (e.g., 5–25 May for winter cereals), created using Cell Statistics—Max.

The specific classification rules implemented in ArcGIS Pro Raster Calculator are as follows:

Winter cereal crops are identified based on the following logical conditions (code 1):

• Con(((“Agri_zone” == 1) & (“PLI_autumn” > 0) & (“PLI_spring” <= 0) &
• (“NDVI_25apr_composite” > “NDVI_25mar_composite”) &
• (“NDVI_15may_composite” > “NDVI_25apr_composite”) &
• (“NDVI_30may_composite” < “NDVI_peak”)), 1, 0)

Here, NDVI_peak represents the composite raster of maximum NDVI values during the typical vegetation peak (5–25 May). PLI_autumn > 0 indicates autumn tillage, while PLI_spring <= 0 ensures the field was not plowed in spring.

Spring cereal crops are identified based on the following conditions (code 2):

• Con(((“Agri_zone” == 1) & (“PLI_spring” > 0) & (“PLI_autumn” <= 0) &
• (“NDVI_05may_composite” > “NDVI_05apr_composite”) &
• (“NDVI_25may_composite” > “NDVI_05may_composite”) &
• (“NDVI_5jun_composite” < “NDVI_peak”)), 2, 0)

Spring tillage is confirmed by PLI_spring > 0, while the absence of autumn tillage is ensured by PLI_autumn <= 0. NDVI_peak reflects the maximum NDVI for 20–30 May.

Safflower is identified based on the following condition (code 3):

• Con(((“Agri_zone” == 1) & (“PLI_spring” > 0) & (“PLI_autumn” <= 0) &
• (“NDVI_25may_composite” > “NDVI_25apr_composite”) &
• (“NDVI_15jun_composite” > “NDVI_25may_composite”) &
• (“NDVI_25jun_composite” < “NDVI_peak”)), 3, 0)

Safflower is identified by a later vegetation peak (10–20 June), spring tillage, and a unique delayed NDVI rise.

Perennial grasses are identified based on the following conditions (code 4):

• Con(((“Agri_zone” == 1) & (“PLI_autumn” < 0) & (“PLI_spring” < 0) &
• (“NDVI_25apr_composite” > “NDVI_25mar_composite”) &
• (“NDVI_15may_composite” > “NDVI_25apr_composite”)), 4, 0)

Perennial grasses exhibit no tillage signals and stable NDVI growth across spring, often with NDVI spikes followed by sharp declines due to harvesting (cutting cycles).

The classification of rainfed crops was based on analyzing temporal NDVI growth profiles, as each crop type demonstrates a unique curve of biomass accumulation during the growing season. For the major crops in the Turkestan region, the following characteristic periods of vegetation onset and peak were defined:

Winter cereals (wheat): Show active early spring growth with peak NDVI between 5 and 15 May in the plains and 15–25 May in foothill zones (depending on thermal conditions). A comparison of NDVI curves between foothill and plain areas is shown in Figure 8.

Spring cereals (barley): Characterized by later vegetation onset and NDVI peaks between 20 and 30 May, lagging behind winter cereals (Figure 9).

Safflower: NDVI curves in both plain and foothill zones show slow biomass accumulation, delayed peak (10–20 June), and rapid post-peak decline during ripening (Figure 10).

Perennial grasses (e.g., alfalfa): Exhibit high, stable NDVI values from March to July, with multiple growth-harvest cycles evident from sharp NDVI drops after biomass accumulation peaks, typically observed in late May and early June (Figure 11).

To enhance the accuracy of crop classification, particularly for species with overlapping phenological profiles or similar vegetation dynamics, the PLI was utilized.

The Plowed Land Index (PLI) was derived from the Tasseled Cap brightness index. The formula for the index represents an original linear combination of values from six spectral bands of Sentinel-2 satellite data: B2, B3, B4, B8, B11, and B12. In constructing this index, each spectral band was weighted and transformed using specific weighting coefficients. The signs in the formula were selected to optimize the detection of plowed land in the foothill and plain zones of the Turkestan region. As a result, a new linear combination was obtained for the detection of plowed land in the study area:

PLI = −0.0037 × b2 − 0.1793 × b3 + 0.5403 × b4 − 0.5585 × b8 − 0.1082 × b11 + 0.3013 × b12

The PLI effectively distinguished between winter and spring crops based on the timing of tillage: winter crops are typically sown after autumn plowing (mid-September to early November), whereas spring crops follow spring tillage (March–April) [15]. This agronomic distinction was integrated into the rule-based classification framework.

The developed classification algorithm combined NDVI temporal profiles with PLI values through logical operators in the Raster Calculator. Additionally, at each classification stage, the Agri_zone.tif mask was applied to restrict the analysis to agricultural land only, excluding non-agricultural areas. This approach significantly streamlined the workflow and enabled full automation without the need for labor-intensive manual digitization of field boundaries. It also ensured scalability for other southeastern regions where similar crops are grown under foothill and plain rainfed conditions.

3. Results

Evaluating the accuracy of the proposed crop classification algorithm is a critical step in validating its robustness and applicability. Unlike traditional pixel-level validation, this study focused on comparing the areas of agricultural crops obtained through automated classification against official statistical records and field survey data. The geographic locations of field surveys used for algorithm calibration and validation are shown in Figure 11. These surveys, conducted across multiple districts of the Turkestan region and supported by key farming enterprises (“Bratya”, “Makulbek”, “Kyzylzhar”, “Karabau”), provided a reliable basis for evaluation.

To ensure a comprehensive assessment and comparison of results, three main data sources on crop areas by district in the Turkestan region for 2022 were utilized:

• Official statistics: Government-reported sown area data for wheat, barley, safflower, and perennial forage crops by district.
• Field boundary-based classification (“mask-based” method): Crop areas derived from a methodology based on detailed digitization of individual field boundaries.
• Rainfed zone classification (“rainfed mask” method): Area estimates for wheat, barley, safflower, and perennial grasses obtained from the developed rule-based classification algorithm applied to the broader Agri_zone.tif mask.

The overall classification results for dominant rainfed crops across the Turkestan region, obtained using the two developed approaches, are presented in Figure 12 and Figure 13.

A visual representation of the key input datasets, the NDVI raster, illustrates the nature of the spectral information used for classification (Figure 8). During automated classification, Sentinel-2 Scene Classification Layers (SCL) were also employed, as illustrated in Figure 12, which enabled efficient masking of clouds and shadows.

3.1. Verification Analysis of RS Data Against Official Statistics

To evaluate the reliability of the developed methods, the crop area estimates derived from RS data were compared with official agricultural statistics at the district level for the SKT region. The relative deviations (errors) were calculated as percentages [53]. The evaluation results—i.e., the deviation percentages (%)—are presented in

Table 2. Deviation of crop area estimates based on remote sensing (field digitization method) compared to official statistics (2022).

Districts	Wheat (%)	Barley (%)	Safflower (%)	Fodder (%)
Arys c.a.	4	27	-	-
Baydibek district	1	2	4	6
Keles district	18	-	-	-
Kazygurt district	0	3	4	3
Ordabasy district	15	3	12	1
Sairam district	16	6	1	6
Saryagash district	11	1	7	3
Sozak district	15	2	2	13
Tolebi district	4	7	3	4
Tulkubas district	6	12	10	4
Mean deviation	9	8	5	5

and

Table 3. Deviation of crop area estimates based on remote sensing (rainfed zone classification method) from official statistics (2022).

Districts	Wheat (%)	Barley (%)	Safflower (%)	Fodder (%)
Arys c.a.	13	27	-	-
Baydibek district	3	27	16	22
Keles district	25	-	-	-
Kazygurt district	2	15	10	14
Ordabasy district	17	-	23	-
Sairam district	5	19	11	15
Saryagash district	3	19	19	12
Sozak district	23	25	-	32
Tolebi district	8	7	18	3
Tulkubas district	8	15	21	7
Mean deviation	11	19	17	15

for two methods: (1) the method using prior field boundary digitization and (2) the method based on the delineation of the rainfed agricultural zone. Calculations were performed only for the records where both the analyzed method and the official statistics contained data. Empty cells (i.e., missing data in either the official statistics or the classification results) were excluded from the error calculation.

The analysis of Table 2 and Table 3 reveals differences in the accuracy of crop area estimation between the two classification methods when compared with official statistical data. It is important to note that average deviation values were calculated only for cells with available data in both sources, excluding empty cells from the calculation.

The results of the classification method based on digitized field boundaries (Table 2) show relatively low average deviations across the territory, ranging from 5% to 9%, demonstrating high reliability and alignment with the actual field boundaries. For individual crop types, the average deviation was approximately 9% for winter wheat, 8% for spring barley, and 5% for safflower and perennial grasses. These findings indicate that the digitized-boundary-based approach yields reliable crop area estimates at the district level and is highly consistent with official statistics.

In contrast, the automated crop recognition method using the rainfed agricultural zone mask, without field digitization, demonstrates somewhat higher deviations at the district level: approximately 11% for winter wheat, 19% for spring barley, 17% for safflower, and 15% for perennial grasses.

Thus, the digitized-boundary method offers better agreement with official statistics for all crop types due to its high spatial detail and the exclusion of non-agricultural areas during the field boundary delineation stage. Although the fully automated classification method using the rainfed zone provides a scalable and efficient solution, its accuracy is slightly lower. The comparative results highlight a clear trade-off between precision and labor intensity. While the digitized-boundary method requires considerable manual effort, the rule-based automated classification method allows for rapid estimation of cultivated crop areas across larger territories with acceptable and operationally useful accuracy.

Observed deviations may be attributed to both methodological differences in data collection (field surveys versus satellite-based assessments) and the presence of fallow or uncultivated lands within the rainfed zone, which could be mistakenly classified as cultivated fields by the algorithm.

3.2. Validation of Classification Accuracy Using Field Survey Data

For a more detailed evaluation and validation of the developed classification methods, field survey data from 2018 and 2022 in the Turkestan Region were utilized. These surveys provided ground-truth information on actual crop distribution.

The field survey dataset includes:

2022: A total of 275 fields, including 25 spring cereals, 35 alfalfa fields, 161 winter wheat fields, and 54 safflower fields.

2018: A total of 335 fields, including 53 spring cereals, 50 alfalfa fields, 77 safflower fields, and 155 winter wheat fields.

In total, 610 fields were surveyed over the two years. These were split into two groups: 310 fields were used for the development and calibration of classification algorithms, while 300 fields were reserved for independent validation of classification accuracy.

Analysis of Classification Accuracy Based on Independent Field Survey Data.

The accuracy assessment based on independent field survey data (

Table 4. Classification accuracy of agricultural crops based on remote sensing data (%) compared with field survey data (2018 and 2022).

Culture	Results of Classification Using the Field Delineation Method	Results of Classification According to the Methodology of Dry Farming
Spring grains	92.5%	80.0%
Alfalfa	90.5%	75.0%
Winter wheat	94.0%	85.0%
Safflower	87.5%	70.0%
Average accuracy	91.1%	77.5%

) supports the general trends previously identified in the comparison with official statistics.

The method based on digitized field boundaries (“by mask”) demonstrates high classification accuracy across all analyzed crop types, with an average accuracy of 91%. Particularly high accuracy was achieved in the classification of winter wheat (94%) and spring cereals (93%). This is expected, as the method operates with pre-vectorized and refined field boundaries, thereby minimizing the impact of “noise” from non-agricultural areas and mixed pixels at field edges.

The developed automated classification method based on the rainfed agriculture zone also yields acceptable results, with an average overall accuracy of 78%. The highest accuracy using this method was observed for winter wheat (85%), due to its distinctive phenological profile. The classification accuracy for spring cereals (80%), alfalfa (75%), and safflower (70%) was slightly lower than the “by mask” method but remained within acceptable levels for regional-scale monitoring and operational mapping. Further improvement in accuracy may be achieved through expert-based refinement and interpretation.

The validation results from field surveys clearly show that the method utilizing digitized field boundaries (“by mask”) outperforms the automated classification method based on the rainfed agriculture zone (“by rainfed area”). This is attributed to fundamental differences in approach: the “by mask” method relies on prior detailed vectorization of fields, which reduces many classification errors, whereas the “by rainfed area” method is fully automated and operates over a broader region, inevitably incorporating some non-agricultural or heterogeneous areas. Nonetheless, the rainfed-zone-based classification method demonstrates sufficient accuracy for large-scale monitoring and crop area estimation, especially where manual digitization is impractical due to the scale and volume of the data. Its key value lies in full automation and the potential for deployment across vast territories with reduced time and resource requirements.

3.3. Validation of Classification Accuracy at the Farm Level

To further evaluate the practical applicability of the developed classification methods, validation was conducted at the level of several key agricultural farms in the SKT region. This analysis aimed to assess how well the classification algorithms perform in real-world farming conditions, which may include unique field structures, crop rotations, and cultivation practices.

Error metrics for each selected farm are presented in

Table 5. Number of classification errors in individual farming enterprises (2022).

Farm	Number of Fields	Errors in the Results of Classification Using the Field Delineation Method (“By Mask”)	Errors in the Results of Classification According to the Method of Dry Farming (“Dry Farming”)
Kyzylzhar	54	0	7
Karabau	94	2	12
Makulbek	7	0	2
Brothers	5	0	0

, showing the number of misclassified fields relative to the total number of observed fields within each farm.

Visualization of classification results for the “Kyzylzhar” farm using both methods is provided in Figure 14 and Figure 15. Figure 14 illustrates the classification output from the method based on digitized field boundaries, showing precise alignment of crop types with digitized polygons. Figure 15 presents the results from the rainfed-zone-based classification method for the same farm, clearly illustrating both correctly classified fields and areas of misclassification or ambiguity within the broader rainfed zone.

The analysis of validation results at the farm level (Table 5) confirms the overall pattern observed in previous comparisons: the field-boundary-based classification method consistently outperforms the rainfed zone–based method in terms of accuracy.

The field-based classification method (“mask-based”) demonstrated a low error rate across the surveyed farms. No classification errors were observed in the “Kyzylzhar,” “Makulbek,” and “Bratya” farms. In the “Karabau” farm, only 2 out of 94 fields (2%) were misclassified. This highlights the high reliability and accuracy of the method when using pre-digitized field boundaries, which helps reduce the influence of external factors and increases confidence in the classification of individual fields.

The rainfed zone–based method, being fully automated, exhibited a higher number of errors at the farm level. Specifically, 7 misclassifications were identified in “Kyzylzhar,” 12 in “Karabau,” and 2 in “Makulbek.” However, it is important to note that in the “Bratya” farm (with 5 fields), both methods showed zero classification errors—likely due to the simplicity and homogeneity of the field structure. The presence of errors in the rainfed zone–based method is primarily due to its reliance on a broad mask that may include not only cultivated areas but also fallow land and vegetatively heterogeneous zones, complicating accurate field-level classification.

These findings indicate that the field-contour-based classification approach is preferable for applications that require high-precision field-level mapping, such as detailed accounting and monitoring within individual farms. In contrast, the rainfed zone–based method serves as an efficient tool for automated large-scale monitoring, where the absolute accuracy of each field is less critical than the speed, scalability, and automation of the process. The errors observed in the latter method represent an acceptable trade-off for timely crop area estimation across vast regions.

4. Discussion

The results of this study evaluating two satellite-based crop classification approaches, one based on digitized field boundaries (“mask-based”) and the other on a generalized rainfed agricultural zone (“rainfed-based”), highlight the respective strengths and limitations of each method, as well as directions for further improvement.

Comparison with official agricultural statistics showed that both methods produced acceptable accuracy levels for estimating crop areas at the district level. The field-boundary method demonstrated higher reliability, as expected, due to its ability to exclude non-agricultural features such as roads, shelterbelts, and infrastructure that could distort satellite data interpretation when using a broader classification mask. However, it should be noted that, for certain crops and districts, the rainfed zone–based method also achieved deviations within acceptable limits for large-scale monitoring. Deviations, especially for fodder crops, may be attributed to the heterogeneous nature of this category, which includes both cultivated perennial grasses and natural pastures or fallow lands, complicating spectral identification via satellite imagery.

Validation using field survey data confirmed these findings, with the field-based method achieving higher classification accuracy (91% on average), compared to 78% for the rainfed zone–based method. The superior performance of the “mask-based” approach stems from the use of vectorized field boundaries, which reduce pixel-level noise and improve classification accuracy. The rainfed zone approach, although less accurate, still demonstrated acceptable performance for a fully automated workflow, making it suitable for rapid assessment across large territories.

Farm-level validation further illustrated these differences in practical applications. The digitized field method yielded almost zero classification errors across the farms analyzed, making it ideal for detailed inventory and management at the individual farm level. The rainfed zone method, despite producing more classification errors on a per-field basis, remains valuable for large-scale automated assessments, especially where detailed field boundary data are unavailable. Visualization of the classification results (Figure 13 and Figure 14, along with maps for the “Kyzylzhar” farm) clearly illustrates the advantage of detailed field digitization in accurately delineating boundaries and crop types compared to the more generalized rainfed mask approach.

In summary, both classification methods offer advantages and are suited for different applications. The field-boundary-based method provides higher accuracy and granularity, making it suitable for tasks requiring precision at the individual field level, particularly in support of farm-level decision-making. However, it involves significant time and resource investments for manual or semi-automated digitization.

In contrast, the rainfed zone–based method offers a fully automated and scalable solution, capable of delivering crop area estimates over large regions with reasonable accuracy. Its key advantages lie in operational efficiency and its applicability in contexts where detailed field mapping is not feasible or cost-effective. This approach is particularly valuable for government agencies and large agribusinesses that require timely statistics and broad-scale trend analysis.

In the analysis of the results, it is important to consider a number of limitations and potential sources of error inherent in our rule-based methodology. One key challenge is the spectral overlap between different agricultural crops, particularly in semi-arid climates. As we have already noted, crops such as safflower and spring cereals can have similar phenological profiles and NDVI dynamics, which complicates their differentiation. Despite the use of additional indicators, such as the Plowed Land Index (PLI), to separate winter and spring crops, it is not possible to completely eliminate these uncertainties.

Another significant source of error can be the heterogeneity of the “forage crops” category, which includes both alfalfa fields and natural pastures. Their similar spectral characteristics and the absence of clear agricultural signs (e.g., plowed) can lead to classification errors. Furthermore, despite the application of a mask to remove clouds and shadows from Sentinel-2 images, residual contamination or an incomplete time series of data during critical phenological periods can also affect the accuracy of the determination. These factors may explain the observed deviations from official statistical data in some areas.

Future research may explore hybrid approaches that combine the strengths of both methods, for example, applying the rainfed zone–based classification for initial estimates, followed by selective refinement using field digitization in critical or problematic areas. Another direction includes integrating additional data sources, such as Synthetic Aperture Radar (SAR) imagery, to improve classification performance under cloud cover and to distinguish spectrally similar crops. Moreover, future research should be dedicated to the comparison of the current rule-based method with machine learning methods such as Random Forest and SVM. Furthermore, the application of the developed methodologies may require additional calibration of NDVI and PLI threshold values when applied in other agro-climatic conditions. Such a comparison would make it possible to assess the relative strengths and weaknesses of each approach and to further validate the effectiveness of the current methodology.

5. Conclusions

This study was dedicated on RS ML applications and comprehensively evaluated two methods for classifying agricultural crops on rainfed lands in the SKT region using Sentinel-2 satellite data: one based on digitized field boundaries (“mask-based”) and the other using a generalized rainfed zone (“rainfed-based”). Both methods demonstrated effectiveness in estimating crop areas for key cultivated crops (wheat, barley, safflower, and perennial grasses), as validated against official agricultural statistics, ground-truth survey data, and farm-level classification accuracy.

The field-digitization method achieved higher classification accuracy (91% on average from ground surveys) and smaller deviations from official statistics, making it well-suited for tasks requiring high precision and spatial detail at the field level.

Meanwhile, the rainfed zone–based approach proved valuable as a fully automated and scalable solution. Although its average accuracy was lower (78%), it provided acceptable performance for large-area crop monitoring with minimal time and resource requirements.

Therefore, the choice of method should align with the specific task: detailed auditing and farm-level management call for field-boundary digitization, while automated large-scale monitoring of crop areas is better served by the rainfed-based approach. The results demonstrate the significant potential of satellite-based monitoring and geospatial technologies for improving the management of the agro-industrial sector and delivering up-to-date information on the condition of rainfed croplands in Kazakhstan.